Parameters
32B
Context Length
128K
Modality
Text
Architecture
Dense
License
Custom Commercial License with Restrictions
Release Date
15 Jan 2024
Knowledge Cutoff
Dec 2023
VRAM requirements for different quantization methods and context sizes
1,024 tokens
Consumer
4x RTX 4090
24GB VRAM
Datacenter
1x NVIDIA A100
80GB VRAM
Apple Silicon
1x Apple M3 Max
128GB VRAM
128,000 tokens
Consumer
4x RTX 4090
24GB VRAM
Datacenter
1x NVIDIA A100
80GB VRAM
Apple Silicon
1x Apple M3 Max
128GB VRAM
No evaluation benchmarks for GLM-4 available.
Overall Rank
-
Coding Rank
-
The GLM-4 32B model is a foundational large language model developed by Z.ai, representing a significant scaling of the General Language Model (GLM) architecture to 32 billion parameters. This model is engineered to balance high-order reasoning capabilities with computational efficiency, serving as a versatile core for advanced agentic applications, complex code generation, and intricate bilingual text processing. It occupies a strategic position within the GLM-4 family, providing the structural complexity necessary for sophisticated linguistic understanding while maintaining a footprint suitable for diverse deployment environments.
Technically, the model utilizes a dense transformer architecture optimized through extensive pre-training on a massive corpus of 15 trillion tokens. This training set includes a substantial proportion of synthetic reasoning data, specifically curated to enhance the model's logical inference and problem-solving skills. The architectural design integrates modern advancements such as Rotary Positional Embeddings (RoPE) and Group Query Attention (GQA), which together facilitate stable performance and efficient inference over a context window of up to 128,000 tokens. To ensure high-quality output, the model undergoes a multi-stage post-training pipeline involving human preference alignment, rejection sampling, and reinforcement learning.
GLM-4 32B is specifically optimized for scenarios requiring structured outputs and autonomous tool interaction. Its performance characteristics make it particularly effective for engineering-grade code generation, precise search-based question answering, and the creation of detailed technical artifacts. The model's refined instruction-following and robust function-calling capabilities enable it to act as the primary engine for intelligent agents that need to plan and execute multi-step tasks across diverse software environments and knowledge domains.
Attention
Attention Structure
Multi-Head Attention
Attention Heads
48
Key-Value Heads
2
Attention Head Dimension
128
Position Embedding
Absolute Position Embedding
RoPE Theta
-
Sliding Window Attention
No
Sliding Window Size
-
Sliding Window Ratio
-
Linear Attention
-
Linear Attention Ratio
-
Normalization
RMS Normalization
Activation Function
SwigLU
Dimensions
Hidden Dimension Size
6,144
Number of Layers
61
FFN Intermediate Size (Dense)
13,696
Multi-Token Prediction Heads
-
Tokenizer
Vocabulary Size
151,552
General Language Models from Z.ai
APX AI
Online